Quantitative Investing (posted 22 Feb 99)

The essence of quantitative investing is crunching numbers to determine whether a proposed investment or portfolio configuration is worthwhile or appropriate. Anything that can go into a digital computer is fair input. And since computers are mostly digital and linear programs are rigid, 'quant' analysis tends to be repetitively structured and rich in reliance on back-testing.

The central themes of quant investing are that history reveals enduring patterns of price behavior, which can be unlocked by statistical techniques; that risk of loss is closely related to volatility, which is related to return; and that management of risk, return, covariance and time frames can be usefully predictable. Part of the attraction of quant tools is cost. All the cost is in setting up the initial framework for analysis, while the incremental cost of each iteration is virtually zero.

Quants often refer to their activity as 'disciplined' with the implication that their approach is less personality prone than others, which rely on qualitative or intuitive judgments to make investment decisions. This is not quite true. Most successful investment managers, whether quant or otherwise, select investment styles consistent with their personalities: they understand the style at its very core, not from the numbers. Indeed, the choice of tools is often artistic, although the repeated application is certainly disciplined or, as some would say, mindless.

Quant investing guru: Robert Arnott

Quant analysis, initially developed in the world of academia, has had an enormous impact on finance practice in the past couple of decades. This story of the 'improbable origins of modern Wall Street' is very well told in Peter Bernstein's book, Capital Ideas.

Several coincident factors led to the widespread popularity of quant investing. First, the availability of academic computer time and willing graduate students boosted empirical financial research in the 1950s and 1960s. Second, a cyclical phenomenon of rising markets led to the use of greater degrees of specificity when the returns seemed to be in application rather than new theories. And third, the graduate students trained in North American universities began to assume decision-making posts around the world, sharing common techniques that encouraged the rapid adoption of new securities and global portfolio construction.

Quant practitioners are comfortable moving back and forth across the borders of the investment and academic worlds. They meet at the same meetings. The well-regarded Q group, for example, has held meetings for a generation blending both communities. Their vocabulary is the same and research standards nearly identical. Academics regularly engage in very well supported research programs funded by investment firms.

Robert Arnott, founder of First Quadrant personifies the quant investing group. He has academic degrees from ????, has written dozens of investment research papers for publication and works tirelessly. His roots are in the University of Chicago and he took his first job as a backroom quant person in a Chicago bank. But he did not stay in the background long. His skill at presenting his own research ideas convincingly to clients and his zeal in work surpassing his peers argued for starting his own firm.

First Quadrant, headquartered in Pasadena, California, has about $26 billion in discretionary assets under management. Its range of investment products is wide since a statistical study of past market behavior reveals the secrets of almost any market. Arnott says: 'We have three elements to our investment philosophy. The first is a pre-disposition toward contrarian investing. The market's don't reward comfort. Secondly, we're quantitative. We try objectively to take emotion out of the investment process. The third element is our use of multi-disciplinary techniques.'

Investment styles are selectively Darwinian in that everyone you talk with is a survivor and can demonstrate that some single mentality has been proven to be correct. Perhaps First Quadrant is best known for tactical asset allocation (TAA). This investment style - which aims scientifically to buy undervalued assets and sell overvalued ones - slides risk up and down according to a pattern, which may be simply stated as deviations from norm. Regression to the mean is the principal tool of proponents of TAA.

Counterpoint

All of us now use quantitative tools though many of these tools, introduced a generation ago and in vogue today, are designed for crude mathematics and limited machines. Dean LeBaron, one of this books' co-authors was an early days enthusiast for quant investing when it was the counterpoint for one-decision, buy and hold investing. As founder of Batterymarch Financial Management, he was an early adopter of academic ideas into investment practice. Now though, when quant investing is the gospel - with a few more agnostics as time goes on - he is less positive.

On the surface, quant analysis is a mechanical application of one or several minds. It is artificial intelligence, which is neither better nor worse than the designer: put some testing device at the end of the process and you have an efficient, automated assembly line. Then you can have a system that is mechanical and learns in a loop, like neural nets and, to a greater degree, genetic algorithms.

But a number of distractions come into the system. One is that it is unclear that the economic hand drives toward profit maximization, total return over whatever time period. The field of behavioral finance is growing (see INVESTOR PSYCHOLOGY), making the study of agents - institutional self-interest - respectable, because it uses different terminology.

Another distraction is the noise in the system, such as partly discontinuous time where markets pulse and do not flow in even time units so there is little serial price correlation. When we model, we make simplifying assumptions that try to capture the essence of the real world. We pretend its exactitude for study, understanding, and, if we are lucky, predictability, assuming it to be linear as if a few units of one variable have an equal effect on another variable over the full range of measurement or estimation. But the real world is distinctly non-linear, like almost everything else in the social sciences.

And we usually ignore the feedback features of markets in which each moment is a unique and unlikely-to-be-repeated event. If market time is in bursts, it may move from regular to chaotic to random. And market study, more than most, suffers from the implications of the Heisenberg principle: that the results we get are often more subject to the identity of the researcher than to the phenomenon being studied.

Another distraction is that there is so much irrelevant information that we do not know the drivers of results. Engineers know how to deal with signal noise, but investment managers do not. Many pride themselves on working only with clean, accurate data, thinking that careless errors are reduced. But techniques to clean the data must be used sparingly and with full knowledge of what is lost and what is gained. The real world is messy and emergent, never like what just happened, which will never happen completely the same again, with all the ramifications. Real data is inherently dirty; model data is usually scrubbed.

The errors introduced by demanding clean data are many. The foremost is time. To get clean data demands that more time is introduced between the last bit of relevant data collected and the time at which it will be used. And time in markets involves, always, the activity of feedback loops that undoes the value of the measurements we have made.

The second error is competitive. If we wait for clean data, knowing that our competitors do as well, we find the answers from that data just when everyone else does, often in the same manner, increasing the likelihood that we arrive at the same, discounted conclusion.

The third error is sample size. To get clean data, we often discard the suspicious measurements, and these may well include the most advantaged (even if startling) and potentially rewarding insights. There are potential events that are not likely to happen but if they do will have big effects on financial markets. Realistic risk assessment would take account of these large low probability events, and the possibility they are more probable than we thought - that there are 'fat tails'.

And finally, there is synthesis where we study smaller and smaller bits-bytes of data, expecting they will retain their characteristics when put back together into a total system.

Most research contains all of these potential flaws. But the overarching question is why quantitative tools that are almost universal like back-testing, but with no basis in academic fact are still used? Why are basic academic principles of statistics, multiple tests, reproducibility, rigorous challenge of results hardly ever observed by quant practitioners? The answer is that they are too busy making money the old-fashioned way, promoting their 'unique' skills. Perhaps learning comes during periods of prosperity and application during periods of adversity.

What passes for investment research is usually back-testing: if I know now what I should have known then, how good the results would be. Even quantitative back-testing is intuitive 'data mining' - determining what patterns exist in a finite sample of numbers. All random number series have a finite beginning and end and patterns can eventually be found - 'torturing the data until it confesses'. Is it any surprise that looking at historical data always produces wonderful suggestions for investment action, but that application of this same process often produces humbling results? Investment strategies based on hindsight often fail.

Our minds, unconsciously perhaps, are conditioned by the immediate past conditions and project variables we would like to include in our models. So it is not at all surprising when we test if these conditioned characteristics are included in our tests, we find that investment results would have been superior. And then we make the leap that they are continuous and we can still capture them because we have not consciously data mined. It would be far better to data mine overtly using the best tools and know what we have done than pretend we did not do it at all.

[Guru response, if any]

Where next?

Just as the failure of Peregrine, the most global and best of the Asian global firms, signals that the Asian investment crisis is very serious, the end of an era, so does the near failure of Long Term Capital Management (LTCM), one of the best of the quantitative firms, signal the end of this phase of interest in quant investing. LTCM engaged in what was supposed to minimal risk investment activities but they made two very low-odds bets, on leverage, both of which hit at the same time.

LTCM made no judgments about the underlying value of the assets it purchased. All its decisions were based on computer models that rigorously excluded any kind of fundamental analysis of companies and currencies. They were only interested in whether the historical relationships between the prices of different assets had changed - not whether they had changed for good reason - on the assumption that you could make money in the future based on what's happened in the past. What is more, they understated the risk of loss by forgetting about large low probability events because their historical data was from a period when these events did not happen.

In doing so, they have reminded us that back-testing and linear investing do not work now at what may be a change point. Instead, what we have reminded ourselves of is that we have been in a generation-long period of acceptance of materials of details about investment by ignoring major market trends - indexing is one of those (see INDEXING), strategy selection by computer is another, trading practices is a third. By emphasizing the details and accepting the theories one could achieve superior results if the trends remained the same.

But we may have had a turn of the trends now. Linear techniques are being disproved and we have not yet determined which of the non-linear techniques will work and how to do it. Research there is too sketchy to apply. We have left one and we have not yet built the foundation of another. It is a time of confusion and sadness to see a firm with the skills of LTCM leave the scene.

But aside from the obvious flaws in conventional management, there are some new approaches coming along. The pattern is remarkably similar to developments in physics at the start of the twentieth century when better laboratory equipment found that Newtonian physics failed to explain experimental results. Something was amiss, and two schools emerged: one trying to restate the old work, better and better, and the other to take radical new departures. And from the latter, we have quantum physics. Perhaps we are at the same juncture in investment management eighty years later.

One of the forward-looking technologies offering promise of better science and better results is drawn directly from physics. It is called complexity, sometimes chaos or adaptive systems, evolutionary dynamics and even artificial intelligence, neural networks and genetic algorithms. The problem with these ideas is that, although they do not have the flaws of conventional practices, they are fuzzy, usually unsupported by repeatable performance attributes, and still undergoing modification. Against the pseudo-precision of old ideas, they seem experimental and flaky. But that is what being on the frontier is all about.

A joke once published in the Journal of Portfolio Management made a very profound point. A new researcher, a freshly minted PhD, came to work at a brokerage firm from academia, and was told that he would do numbers, problems involving addition, subtraction, multiplication and division - but not to worry, division was not used very much. The point of that story was that what passes for quantitative analysis in the investment arena is frequently static, simplistic and does not meet the rigors of statistical tests that can be used with today's instruments and today's demands.

The new tests are dynamic, often involving complexity analysis, often involving destructive statistical testing to determine the limits, are very demanding of time scales and frequently are so large and so demanding, such as high-frequency analysis, that they can be done only on supercomputers. In today's investment world, quantitative analysis must be the very best, not the most used.

Read on

In print

Peter Bernstein, Capital Ideas: The Improbable Origins of Modern Wall Street

Online

www.santafe.edu - website of the Santa Fe Institute
mitpress.mit.edu/SNDE - Studies in Nonlinear Dynamics & Econometrics features the best work of academic quants